search

samet-akcay / ganomaly

GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training
MIT License
853 stars 213 forks source link

Test problem #97

Open qianxiao111 opened 1 year ago

qianxiao111 commented 1 year ago

What is it about testing the same abnomal image with a trained model that gives different results(AUC) each time? How can I solve this problem? And, When a normal sample were put into the abnomal folder for test, the test results is also bad.

caiyu6666 commented 1 year ago

I found that in the testing code (lib/model.py, line 175), the netowork is not turned into the evaluation mode (missing self.netg.eval() and self.netd.eval()). As a result, the model will update the mean and var in BatchNorm2d layers.

I think that is the reason why the testing results change in each testing time.

Additionally, I found that the missing of net.eval() will cause higher performance, as the model can learn something from the testing set. So I think this bug can lead to unconvincing results.

tobybreckon commented 1 year ago

See discussion under issue https://github.com/samet-akcay/ganomaly/issues/83